Cocoa audits were falsified in one case in four. Hiding the pass mark cut it in half

Researchers led from ETH Zurich ran a field experiment across 407 cocoa farms in Cote d'Ivoire and found that one in four monitoring entries had been manipulated to meet a sustainability program's requirements ^[1][2][3]. When auditors did not know the target value they were measuring against, the rate of fraudulent alterations halved ^[4], which makes this one of the rare integrity problems with a fix that costs nothing but a change to a form.

The setting was a shade-tree planting program run by an international cocoa buyer and exporter, in which smallholders were invited to plant trees to protect long-term cocoa health and raise biodiversity ^[5][6]. The researchers changed one thing in the monitoring system: farmers were rewarded for reporting their tree numbers accurately ^[7]. Auditors then visited, counted trees, and entered the count by smartphone into a system that immediately told them whether their count matched the farmer's figure and therefore whether the farmer had earned the reward ^[8]. Auditors could also go back and correct entries after the fact ^[9].

The pattern in those corrections is the finding. Where the farmer's count was confirmed on the first attempt, retroactive changes appeared in 6% to 7% of cases, according to lead author Federico Cammelli ^[10]. Where it was not confirmed, changes appeared in 33% of cases ^[11], roughly five times as often ^[1]. Criteria that had no bearing on the audit or the reward were not altered at all ^[12]. Cammelli's reading is that auditors "very specifically adjusted only the data relevant to passing the audit" ^[13]. This is not measurement noise or fat fingers; it is directional.

The incentive structure explains why. Auditors were paid nothing for accuracy, but had reasons to want farms to pass, including preserving working relationships with farmers or sharing the reward with them ^[14]. Cammelli's broader conclusion after several years studying cocoa certification is that there are conflicts of interest along the whole value chain and no incentive to have farm-level data independently verified ^[15].

So the experiment split farms into two arms, one in which auditors knew the parameters used to decide whether a farmer passed ^[16]. Concealing those values cut false reporting by half ^[4]. Put in operator terms, the manipulable ingredient was not the auditor's ethics but the interface: a real-time pass/fail indicator plus an edit-after-submission function tells the data collector exactly which number to move and by how much. Halving the fraud rate meant removing the signal, not adding supervision ^[2].

The study is published in Science ^[17], with co-authors including Johan Six of ETH Zurich and Rachael Garrett of the University of Cambridge ^[18]. Its relevance is not confined to chocolate wrappers carrying Rainforest Alliance or Fairtrade marks ^[19]. The publisher frames the findings as inputs for companies, public authorities and certification bodies building sustainable supply chains and implementing the EU Deforestation Regulation ^[20], where the same architecture is being rebuilt at scale: field data, mobile capture, a threshold, and a party with an interest in the answer.

Worth watching: whether any certification body actually blinds thresholds in its audit apps, and whether EUDR due diligence tooling ships with compliance status visible at the point of entry. Also worth noting that this is one study of one program in one country, and the 25% figure is a rate of manipulated entries in a monitoring dataset the researchers had deliberately given farmers a reason to game ^[3][7].